Investigate Helion for kernel authoring/tuning

[shimizust]

🚀 The feature, motivation and pitch

Summary

Helion is PyTorch's new DSL that compiles to Triton and has powerful autotuning capabilities. It could make Liger kernels easier to write and more optimized across GPU architectures/problem sizes.

We could tune for some common architectures (e.g. H100/B200) and fallback to some sensible defaults otherwise.

Potential risks or issues

• Helion is experimental (bugs, API changes)
• Requires additional dependency from users, torch 2.9
• Requires pre-tuning kernels across multiple architectures and shipping several configs, as tuning can take a long time (~10 min)

Alternatives

No response

Additional context

No response

[Tcc0403]

I can work on a simple version of FLCE

[shimizust]

@Tcc0403 that would be great! I can also do some exploration

[Tcc0403]

FLCE on #928